The use of data mining and adaptive learning is a very controversial issue among the algorithmic trading community in the financial world. The reason for the mistrust in the techniques arrives from some very well-known problems: overfitting to training data and insufficient support for the derived models. In this paper, we present a new element to the use of some classic data mining and adaptive learning algorithms: a set of objective distance measurements that track the similarity between the prediction model and the actual system. We use historical market data to develop algorithms, and investigate the correlation between prediction accuracy of the models and their distance measurements. We find that this tracking could allow investors to discard stale models earlier, thus decreasing losses.
展开▼
